perf: Fix indexes on singleevent table [DHIS2-20891]

[enricocolasante]

Before (6 single-column indexes):

• (status, occurreddate)
• (occurreddate)
• (attributeoptioncomboid)
• (programstageid)
• (organisationunitid)
• (deleted, assigneduserid)

After (3 composite indexes, all leading with programstageid):

• (programstageid, occurreddate)
• (programstageid, organisationunitid)
• (programstageid, assigneduserid)

Following up on #22971, the key insight is that single event queries always filter by program stage — this is now explicitly enforced in JdbcSingleEventStore where ev.programstageid = :programstageid is added as a mandatory filter. Since programstageid is always in the WHERE clause, making it the leading column in every composite index means:

• Every query can use an index scan from the first column. A B-tree index is only efficiently usable when you filter on a left prefix. The old single-column indexes on occurreddate, organisationunitid, or assigneduserid couldn't be combined with the programstageid filter in a single index scan — Postgres would have to do a bitmap AND of two separate index scans or fall back to a sequential scan.
• Dramatically reduces the working set. The programstageid filter is highly selective (narrows to one program's events). The second column then further narrows within that subset, making range scans (e.g. occurreddate for date range queries) and lookups (e.g.organisationunitid for org unit filtering) very efficient.
• Fewer indexes = less write overhead. Going from 6 indexes to 3 reduces the cost of every INSERT/UPDATE/DELETE on the singleevent table.
• The dropped (status, occurreddate) and (attributeoptioncomboid) indexes were either unused or covered by other access paths.

Performance

Compared against 0de04dd (master at time of PR).

baseline:  dhis2/core-pr:22971
candidate: dhis2/core-pr:22977

• Smoke run

Request	Base p95	Cand p95	Diff
Get first enrollment	11ms	10ms	-9%
Get first event	21ms	20ms	-5%
Get first event from enrollment	24ms	25ms	+4%
Get first page of TEs of program ur1Edk5Oe2n	28ms	28ms	+0%
Get first tracked entity	36ms	38ms	+6%
Get relationships for first event	4ms	4ms	+0%
Get relationships for first tracked entity	6ms	5ms	-17%
Get TEs with enrollment status	33ms	34ms	+3%
Get tracked entities from events	7ms	6ms	-14%
Go to first page of program VBqh0ynB2wv	14ms	13ms	-7%
Go to second page of program VBqh0ynB2wv	15ms	13ms	-13%
Not found TE by name with like operator	11ms	11ms	+0%
Not found TE by national id with eq operator	9ms	8ms	-11%
Search events by program stage	10ms	9ms	-10%
Search single events assigned to any user in program VBqh0ynB2wv	8ms	7ms	-12%
Search single events in date interval in program VBqh0ynB2wv	31ms	29ms	-6%
Search single events not assigned to a user in program VBqh0ynB2wv	14ms	12ms	-14%
Search TE by name with like operator	27ms	28ms	+4%
Search TE by national id with eq operator	10ms	9ms	-10%

• Load run

Request	Base p95	Cand p95	Diff
Get first enrollment	27ms	28ms	+4%
Get first event	41ms	40ms	-2%
Get first event from enrollment	76ms	76ms	+0%
Get first page of TEs of program ur1Edk5Oe2n	47ms	47ms	+0%
Get first tracked entity	55ms	56ms	+2%
Get relationships for first event	4ms	4ms	+0%
Get relationships for first tracked entity	5ms	5ms	+0%
Get TEs with enrollment status	54ms	54ms	+0%
Get tracked entities from events	13ms	13ms	+0%
Go to first page of program VBqh0ynB2wv	11ms	11ms	+0%
Go to second page of program VBqh0ynB2wv	11ms	11ms	+0%
Not found TE by name with like operator	11ms	12ms	+9%
Not found TE by national id with eq operator	8ms	8ms	+0%
Search events by program stage	20ms	19ms	-5%
Search single events assigned to any user in program VBqh0ynB2wv	7ms	7ms	+0%
Search single events in date interval in program VBqh0ynB2wv	11ms	11ms	+0%
Search single events not assigned to a user in program VBqh0ynB2wv	11ms	11ms	+0%
Search TE by name with like operator	43ms	42ms	-2%
Search TE by national id with eq operator	9ms	9ms	+0%

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[teleivo]

Dramatically reduces the working set. The programstageid filter is highly selective (narrows to one program's events)
The selectivity depends on the actual data.

Why don't we see more of an effect in any of the response times? 1-2ms is not that dramatic 😅 Although definitely good from the write end to remove the overhead!

[teleivo]

Why this date and not (also) others? I mean it already had an index but why this date vs another?